https://github.com/aiwaves-cn/RecurrentGPT

Official Code for Paper: RecurrentGPT: Interactive Generation of (Arbitrarily) Long Text
https://github.com/aiwaves-cn/RecurrentGPT

Last synced: about 1 month ago
JSON representation

Official Code for Paper: RecurrentGPT: Interactive Generation of (Arbitrarily) Long Text

• Host: GitHub
• URL: https://github.com/aiwaves-cn/RecurrentGPT
• Owner: aiwaves-cn
• License: gpl-3.0
• Created: 2023-05-22T15:43:19.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2024-05-15T03:35:35.000Z (7 months ago)
• Last Synced: 2024-11-09T11:37:04.329Z (about 1 month ago)
• Language: Python
• Homepage:
• Size: 2.41 MB
• Stars: 971
• Watchers: 12
• Forks: 155
• Open Issues: 13
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• StarryDivineSky - aiwaves-cn/RecurrentGPT

README

        
# RecurrentGPT

[📄 Paper] | [🤗 Demo - Writing Assistant] | [🤗 Demo - Interactive Fiction] |  [📺 Video] | [🔥 Discord] 



## Framework Illustration







> RecurrentGPT replaces the vectorized elements (i.e., cell state, hidden state, input, and output) in a Long-short Term Memory RNN (LSTM) with natural language (i.e., paragraphs of texts), and simulates the recurrence mechanism with prompt engineering. 

> At each timestep t, RecurrentGPT receives a paragraph of text and a brief plan of the next paragraph, which are both generated in step t − 1. It then attends to the long-term memory, which contains the summaries of all previously generated paragraphs and can be stored on hard drives, and relevant paragraphs can be retrieved with semantic search.

> RecurrentGPT also maintains a short-term memory that summarizes key information within recent timesteps in natural language and is updated at each time step. RecurrentGPT combines all aforementioned inputs in a prompt and asks the backbone LLM to generate a new paragraph, a short plan for the next paragraph, and updates the long-short term memory by rewriting the short-term memory and appending the summary of the output paragraph to the long-term memory. 

### Example







## Deployment

You can change the configurations given in the `recurrent.sh` script

```yaml

iteration: 10                       #(int) the number of rounds you would like it to roll.

outfile: response.txt               #(str) the output file path.

init_prompt: init_prompt.json       #(str) the path to the prompt used for initialization.

topic: Aliens                       #(str) the topic that you wish your novel is about.

type: science-fiction               #(str) the type of novel you would like to write.

```

Then after specify your `OPENAI_API_KEY` in the `recurrent.sh` file, you can run

```

sh recurrent.sh

```

NOTE: If your local internet is not allowed to access OpenAI's API, you might need to first export your HTTP proxy in the `recurrent.sh` file as well.

```

export http_proxy='your_proxy'

```

## Showcases

### Prompt Engineering







### Iterations







> Human writer starts by choosing the topic he/she wants to write and writes a short paragraph describing the background and the outline of the book. Then RECURRENTGPT automatically generates the first paragraphs and provides a few possible options for the writer to continue the story. The writer may select one from them and edit it if needed. He or she can also write a short plan for the next few paragraphs by him/herself if generated plans are all inappropriate, which makes human-AI co-writing process more flexible

## Web demo

You can directly use our online demo at:

https://www.aiwaves.org/recurrentgpt  and https://www.aiwaves.org/interactivefiction

Or you can run it on your local machine by editing the OPENAI_API_KEY and OPENAI_Proxy in utils.py and then run:

```

python gradio_server.py

```

![web-demo](resources/web_demo.png)

## Use customized LLMs for local deployment

Please refer to https://github.com/jackaduma/Recurrent-LLM to use opensource LLMs for local deployment. Many thanks to @jackaduma 

## Citation

```angular2

@misc{zhou2023recurrentgpt,

      title={RecurrentGPT: Interactive Generation of (Arbitrarily) Long Text}, 

      author={Wangchunshu Zhou and Yuchen Eleanor Jiang and Peng Cui and Tiannan Wang and Zhenxin Xiao and Yifan Hou and Ryan Cotterell and Mrinmaya Sachan},

      year={2023},

      eprint={2305.13304},

      archivePrefix={arXiv},

      primaryClass={cs.CL}

}

```